CN110263866A - A kind of power consumer load setting prediction technique based on deep learning - Google Patents
A kind of power consumer load setting prediction technique based on deep learning Download PDFInfo
- Publication number
- CN110263866A CN110263866A CN201910550680.7A CN201910550680A CN110263866A CN 110263866 A CN110263866 A CN 110263866A CN 201910550680 A CN201910550680 A CN 201910550680A CN 110263866 A CN110263866 A CN 110263866A
- Authority
- CN
- China
- Prior art keywords
- load
- value
- prediction
- point
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 69
- 238000013135 deep learning Methods 0.000 title claims abstract description 32
- 238000013528 artificial neural network Methods 0.000 claims abstract description 18
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 15
- 230000000306 recurrent effect Effects 0.000 claims abstract description 11
- 230000002159 abnormal effect Effects 0.000 claims description 47
- 238000012360 testing method Methods 0.000 claims description 24
- 238000012549 training Methods 0.000 claims description 22
- 230000006870 function Effects 0.000 claims description 18
- 238000005457 optimization Methods 0.000 claims description 12
- 239000013256 coordination polymer Substances 0.000 claims description 11
- 238000000354 decomposition reaction Methods 0.000 claims description 9
- 230000003321 amplification Effects 0.000 claims description 8
- 238000011156 evaluation Methods 0.000 claims description 8
- 238000003199 nucleic acid amplification method Methods 0.000 claims description 8
- 230000005856 abnormality Effects 0.000 claims description 7
- 230000001186 cumulative effect Effects 0.000 claims description 7
- 238000012937 correction Methods 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 5
- 238000013480 data collection Methods 0.000 claims description 4
- 238000011160 research Methods 0.000 claims description 4
- 230000009977 dual effect Effects 0.000 claims description 3
- 240000002853 Nelumbo nucifera Species 0.000 claims description 2
- 235000006508 Nelumbo nucifera Nutrition 0.000 claims description 2
- 235000006510 Nelumbo pentapetala Nutrition 0.000 claims description 2
- 229910052729 chemical element Inorganic materials 0.000 claims description 2
- 230000001537 neural effect Effects 0.000 claims 1
- 238000012706 support-vector machine Methods 0.000 claims 1
- 238000010801 machine learning Methods 0.000 abstract description 2
- 238000007781 pre-processing Methods 0.000 abstract 1
- 230000000694 effects Effects 0.000 description 13
- 230000005611 electricity Effects 0.000 description 6
- 241001269238 Data Species 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000007774 longterm Effects 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 241000208340 Araliaceae Species 0.000 description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 235000008434 ginseng Nutrition 0.000 description 2
- 230000009885 systemic effect Effects 0.000 description 2
- 230000002547 anomalous effect Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000012790 confirmation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000007717 exclusion Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000000446 fuel Substances 0.000 description 1
- ZZUFCTLCJUWOSV-UHFFFAOYSA-N furosemide Chemical compound C1=C(Cl)C(S(=O)(=O)N)=CC(C(O)=O)=C1NCC1=CC=CO1 ZZUFCTLCJUWOSV-UHFFFAOYSA-N 0.000 description 1
- 230000030279 gene silencing Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 210000003739 neck Anatomy 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000003908 quality control method Methods 0.000 description 1
- 238000011158 quantitative evaluation Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
Abstract
The power consumer load setting prediction technique based on deep learning that the invention discloses a kind of, comprising the following steps: (1) establish large user's historical load data pretreated model;(2) the load point prediction model based on LSTM time recurrent neural network is established;(3) the load setting prediction algorithm of point prediction value scaling coefficient is used.Through the above way, the present invention is to carry out Preprocessing to single user's historical data by establishing the customer charge pretreated model based on state vector machine method, according to treated, historical data uses LSTM machine learning method to find in the prediction model for reducing customer charge prediction error to greatest extent, the load setting prediction of single user is carried out with point prediction value scaling coefficient load interval prediction algorithm, accurate load setting prediction can be carried out to the load of the single power consumer with strong stochastic volatility, traditional method is substantially better than in the prediction accuracy of customer charge.
Description
Technical field
The present invention relates to power system automatic fields, more particularly to a kind of power consumer load based on deep learning
Interval prediction method.
Background technique
The complication system that electric system is made of power plant, transmission line of electricity, distribution system and load, electric system
Economical operation is to provide electric power to user with least cost under conditions of meeting safe and reliable, and load prediction is as energy pipe
Reason system (EMS) and the important component of electricity market operational management, prediction result and electric system it is safe, economical
It runs closely related.
Load prediction is divided into ultra-short term, short-term, medium and long term prediction, super short period load according to the unusual of target
Prediction refers to the load prediction in one hour following, is mainly used for utility power quality control, security monitoring, prevention and emergent control
Deng;Short-term load forecasting refers to load prediction in following one day to several days, is mainly used for Optimization of Unit Commitment, economic trend control
System, hydro thermal coordination etc.;Medium term load forecasting refers to the load prediction several months ahead of time implemented by 1 year, is mainly used for reservoir tune
Degree, fuel planning and unit maintenance etc.;Long term load forecasting refers to the load prediction that the several years implement in advance, is mainly used for power grid
Reconstruction, the perspective long-term plan of system, new power plant invest to build.
Current existing load prediction technology and method are the region load for entirety mostly, and to single load user
Prediction it is rarely found, with electric power demand side reform and electricity market propulsion, to the detailed predicting of large user Systemic Burden
It is particularly important;However, the load prediction of user class and traditional regional, system-level load prediction have larger difference, it is main
Be embodied in: the load of regional systems grade is the comprehensive effect of a large amount of Systemic Burdens, since there are certain journeys for fluctuation between individual
The counteracting of degree, so that the fluctuation that regional systems stage load is presented is not obvious;And it is bent to observe a large amount of large user's daily load
Line is it can be found that the production process as specific to its industry industry or business activity rule, and the load curve of user class is at it
On the basis of individualized feature, there are biggish stochastic volatilities, so, the existing prediction technique for region load is being retouched
Customer charge is predicted and is not suitable in terms of stating stochastic volatility, therefore, studies and is suitble to the fine section of user's stage load pre-
Survey method is imperative.
Summary of the invention
The power consumer load setting prediction based on deep learning that the invention mainly solves the technical problem of providing a kind of
Method.
In order to solve the above technical problems, one technical scheme adopted by the invention is that:
A kind of power consumer load setting prediction technique based on deep learning is provided, comprising the following steps:
(1) identification and repairing of large user's historical load data pretreated model for abnormal data are established;
(2) the load point prediction model based on LSTM time recurrent neural network is established;
(3) the load setting prediction algorithm of point prediction value scaling coefficient is used.
In a preferred embodiment of the present invention, the abnormal data includes dominant abnormal data and recessive abnormal data.
In a preferred embodiment of the present invention, every 15 points in AMI system the identification of the dominant abnormal data: are checked
Clock any the discovery of large user's load data there are load data sometime point or certain continuous moment points lack the case where,
It needs to fill up above-mentioned record with zero, and recording exceptional mark;For already present data point in AMI system, will deposit
It picks out to come in the extreme abnormal conditions of numerical value, and recording exceptional mark;If the abnormality mark of above-mentioned record is present in k consecutive hours
In quarter, then the 96 point load values of this day is deleted from sample database, otherwise repaired;
The repairing step of the dominant abnormal data:
A) it on the basis of day where moment point to be modified, respectively forwardly and backward finds m days, if the m days in the future loads
Point has a dominant abnormality mark and uncorrected, postpones backward one day;Find out respectively 2m days and moment point to be modified where day day
Equal load, it is contemplated that the difference on working day and day off, the similar day carried out to above-mentioned 2m+1 days based on average daily load are classified, asked
Out with the same type day of day where moment point to be modified;
B) same type day is constructed respectively comprising time window composed by n before the amendment moment continuous moment points, from load
Two aspects of size magnitude and load fluctuation trend, when carrying out this 2m time window curve with where abnormal moment point respectively
Between window curve similarity research, be based primarily upon the method for Euclidean distance similarity and the method based on cosine similarity, find
At magnitude and the aspect of fluctuation tendency two with abnormal point where the more similar time window of time window curve as supporting vector
The training sample of machine SVM.Based on these samples, the fitting based on SVM is carried out to the load at moment where abnormal load data point
Training, later using the data sequence before abnormal load data to be modified in time window as being input to trained SVM mould
In type, the correction value of dominant abnormal data is obtained;
For given load data collectionIt is fitted with such as drag:
F (x)=(wx)+b (1)
Wherein, the number for the load similar day (the similar day of load form) that wherein n is taken by load sample data set, yi
For the load of object time point, xiThe vector for being d for dimension, d time point is corresponding before numerical value is object time point
Load;
W is real constant vector, w ∈ Rd, b is real constant, and the numerical value of b ∈ R, w and b are based on sample data set (xi,yi)n
Fitting obtains, and introduces the first slack variable ξiWith the second slack variableConstruct following optimization problem:
Wherein, constant C is penalty coefficient, and ε is prescribed skew value;
The optimization problem of its dual spaces are as follows:
Wherein, the first Lagrange multiplier αiWith the 2nd Lagrange multiplier αi *;
This problem is solved, the first optimal Lagrange multiplier α is obtainediWith the 2nd Lagrange multiplier αi *, to obtain
Fitting function
Wherein, K (x, xi)=(xxi+ 1) the load correction value repaired needed for, gained fitting function f (x) is here.
In a preferred embodiment of the present invention, using wavelet decomposition and the method for wavelet reconstruction in customer charge sequence
Recessive abnormal data the step of being recognized: wavelet transformation is carried out to 96 point data of daily load first, chooses wavelet basis
Signal is carried out 4 layers of decomposition by db4, low frequency part after decomposition correspond to signal trend part be a relative smooth song
Line, corresponding wavelet coefficient are wk(k=1,2 ..., N) carries out the denoising based on soft-threshold to high frequency section signal,
It is taken based on the weighted average threshold function table of soft and hard threshold function:
Wherein, weighted factorT is threshold value thresholding, wkFor kth time wavelet systems
Number, using the method for the VisuShrink of fixed threshold threshold criterion come threshold value thresholdingWherein σ is to make an uproar
The variance of sound, further according to the wavelet coefficient after soft-threshold de-noisingWavelet reconstruction after de-noising, root are carried out to high frequency section signal
Rough error position is judged according to the extreme point of signal after high frequency section signal noise silencing, and is excluded thick as caused by fluctuation situation
Almost, if being confirmed as recessive exceptional value, the modified method of dominant exceptional value is taken to be modified.
In a preferred embodiment of the present invention, according to large user's historical load data pretreated model, from electricity consumption
The historical load sequence that user every 15 minutes intervals are extracted in information acquisition system forms raw data set, to the daily load of user
96 point datas do prediction a few days ago and obtain customer charge predicted value;It is the instruction of neural network based on actual power consumer load data
Practice collection, test obtains optimal time recurrent neural network LSTM hidden layer configuration;With the equal of customer charge predicted value and actual value
Square error updates each section weight of LSTM model with the minimum optimization aim of loss parameter as loss parameter, obtains most
Good load forecasting model.
In a preferred embodiment of the present invention, the time recurrent neural network LSTM hidden layer configuration includes:
If unit exports h, input data x, g are LSTM unit output quantity, and i is the output quantity of input gate, and f is that forgetting door is defeated
Out, o is out gate output quantity, and c is memory unit output quantity, and h is entire LSTM unit output quantity, bg、bi、bf、boIt is reference
The update of base value, the unit of LSTM time recurrent neural network is as follows:
gt=tanh (xtwxg+ht-1whg+bg) (6)
it=sigmoid (xtwxi+ht-1whi+ct-1wci+bi) (7)
ft=sigmoid (xtwxf+ht-1whf+ct-1wcf+bf) (8)
ot=sigmoid (xtwxo+ht-1who+bo) (9)
Extracted from power information acquisition system user nearly 2 years every 15 minutes load sequences, form initial data
Collection, concentrates all adjacent two days electrical load data to be combined into a record initial data, the previous day is as LSTM network
Input constitute data set D one day after as the label of network or output, enable the record for accounting for 90% in data set D as instruction
Practice collection M, accounts for 10% record as test set N, remember
The record number of training set is m, and the record number of test set is n;
LSTM model is using the mean square error of predicted value and actual value as loss parameter, with the minimum optimization mesh of loss parameter
Mark carrys out each section weight of more new model, defines the consensus forecast deviation of all moment points:
Wherein real is the true value in test set sometime, and pred is the predicted value of its corresponding LSTM model, p=
96。
In a preferred embodiment of the present invention, according to the load point prediction model to as caused by uncertain factor
The mobility scale of prediction load quantified, the forecast interval that Lower and upper bounds determine is provided, so that actual load observation is with one
Fixed expected probability is fallen in the section.
In a preferred embodiment of the present invention, the evaluation index packet of the variation range of the prediction load of the forecast interval
Include section coverage rate χCP, mean breadth percentage χMWPWith cumulative departure χAD;
Section coverage rate χCPBe actual value fall in by the upper bound, lower bound envelope forecast interval in probability, actual value
realijIt is fallen in the forecast interval of building with the probability not less than specified confidence level, it may be assumed that
P(realij∈[L(predij),U(predij)])≥μ (13)
Wherein, L (predij) and U (predij) it is by point prediction value pred respectivelyijThe lower bound of obtained forecast interval and upper
Boundary, μ are that specified confidence interval is horizontal, and the corresponding section coverage rate of j-th of moment point is defined as follows:
Wherein,
Mean breadth percentage χMWPThe average percent that forecast interval width accounts for true value, j-th of moment point pair are measured
The mean breadth percentage answered is defined as follows:
Cumulative departure χADEmbody the degree that actual load observation deviates forecast interval, the accumulation of j-th of moment point is inclined
Difference are as follows:
Wherein,
In a preferred embodiment of the present invention, the Satisfaction index χ of the forecast intervalPISIIn j-th of moment point are as follows:
Wherein, η is to χCPjThe penalty coefficient of value, χPISIjThe upper bound corresponding to minimum value and lower bound are finally to choose
Interval prediction is as a result, work as χMWPjAnd χCPjWhen identical, according to χADjValue is selected, by the smallest χADjThe corresponding upper bound of value and
Lower bound is as the interval prediction result finally chosen.
In a preferred embodiment of the present invention, the load setting prediction algorithm is using each moment as basic calculating list
Member carries out interval prediction to each moment respectively, and the j moment shares m predicted value in training set and true Value Data is corresponding, will
The upper bound and lower bound of the result that ratio factor alpha is amplified respectively and reduced to predicted value and β is obtained as forecast interval:
It determining amplification and reduces the value of ratio factor alpha and β, precision k is the precision of α and β, it enables:
α’And β’For the possibility value of α and β, Ceil function is above to enter bracket function, carries out calculated result to entire training set
Duplicate removal can obtain the value range of α and β, choose α and β conduct corresponding when the Satisfaction index minimum of forecast interval
The amplification of the moment point and diminution proportionality coefficient, which is used in test set, obtains the upper bound and lower bound according to predicted value,
By the relationship of actual value and the upper bound, lower bound, the areal coverage and average width percentage of test set are calculated, verifying section is pre-
Survey result.
The beneficial effects of the present invention are: providing a kind of power consumer load setting prediction technique based on deep learning, lead to
Customer charge historical data pretreated model of the foundation based on state vector machine method is crossed to carry out in advance single user's historical data
Processing analysis, according to treated, single user's historical data uses the method searching of LSTM machine learning to be intended to the maximum extent
The prediction model of customer charge prediction error is reduced, is carried out with point prediction value scaling coefficient load interval prediction algorithm single
The load setting of user is predicted, can carry out accurate loading zone to the load of the single power consumer with strong stochastic volatility
Between predict, traditional method is substantially better than in the prediction accuracy of customer charge.
Detailed description of the invention
To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment
Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for
For those of ordinary skill in the art, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing, in which:
Fig. 1 is the LSTM of power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
Cellular construction figure;
Fig. 2 is the section of power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
Prediction algorithm explanatory diagram;
Fig. 3 (a) is power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
Correct day before yesterday load data figure;
Fig. 3 (b) is power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
Daily load datagram after amendment;
Fig. 4 (a) is power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
At the time of user two days 58,59,62 be missing point daily load curve figure;
Fig. 4 (b) is power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
At the time of user two days 64,89 be missing point daily load curve figure;
Fig. 5 (a) is power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
The revised daily load curve figure of user's missing values one day;
Fig. 5 (b) is power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
Several days load charts in front and back of day where the revised daily load curve of user's missing values one day and abnormal load point;
Fig. 6 is the LSTM of power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
(H1=100, H2=0) model point prediction result figure a few days ago;
Fig. 7 be power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning based on
Interval prediction result figure of the interval prediction algorithm of LSTM point prediction under different confidence levels;
Fig. 8 is the dominant of power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
The Technology Roadmap of abnormal data repairing;
Fig. 9 is power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning containing single
The typical RNN of hidden layer schemes;
Figure 10 is the exhibition of power consumer load setting prediction technique one preferred embodiment of the invention based on deep learning
RNN figure after opening.
Specific embodiment
The technical scheme in the embodiments of the invention will be clearly and completely described below, it is clear that described implementation
Example is only a part of the embodiments of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, this field is common
Technical staff's all other embodiment obtained without making creative work belongs to the model that the present invention protects
It encloses.
Fig. 1 to Figure 10 is please referred to, the embodiment of the present invention includes:
A kind of power consumer load setting prediction technique based on deep learning, comprising the following steps:
(1) large user's historical load data pretreated model is established
The pretreatment of large user's historical load data refers mainly to the identification and amendment of exception history load data, abnormal number
According to: record missing values, beyond power zero, meter zero point brought by user transformers load limit, power failure or communicating interrupt
Drift about the smaller value etc. caused, and at the time of exception history load data occurs and magnitude has very strong randomness, faces magnanimity
User's history load data, by the mode of artificial eye go distinguish seemed unable to do what one wishes, research be based on artificial intelligence side
The discrimination of the exception history load data of method and amendment are imperative.
From load form, the transient fluctuation of customer charge is likely to be abnormal data, it is also possible to be production process
Intrinsic impact, and found from the analysis of existing subscriber's load data form, the daily load form of many users is when adjacent
Between there is very strong fluctuation between point, and between day and day due to production procedure etc., time point is corresponding with fluctuation
Without good consistency between relationship;Form is fluctuated in face of such complicated customer charge, studies distinguishing automatically for abnormal data
Know and the repairing technique mainly exhibition in terms of the identification of dominant abnormal data and repairing, the identification of recessive abnormal data and amendment two
It opens.
(1) identification and repairing of dominant abnormal data
Check every 15 minutes any large user's load datas in AMI system it can be found that there are load datas at certain
The case where a moment point or certain continuous moment points lack, i.e., record without this, firstly, it is necessary to by above-mentioned record with zero
Value is filled up, and recording exceptional mark, is repaired according still further to the method for repairing and mending of following abnormal datas;On the other hand, for
Already present data point in AMI system can have the extreme abnormal conditions of following numerical value, such as: beyond user transformers load limit,
The smaller value etc. that power zero brought by power failure or communicating interrupt, meter null offset cause, these types of extreme case is distinguished
Knowledge comes out, and recording exceptional mark, if the abnormality mark of above-mentioned record was present in k continuous moment, by 96 points of this day
Load value is deleted from sample database, is otherwise repaired according to following method for repairing and mending, and step is repaired:
A) it on the basis of day where moment point to be modified, respectively forwardly and backward finds m days, if the m days in the future loads
Point has a dominant abnormality mark and uncorrected, postpones backward one day, to avoid the uncorrected point of dominant exception to based on average negative
The influence of lotus similar day classification;Find out respectively 2m days and moment point to be modified where day average daily load, it is contemplated that working day with
The difference on day off, the similar day carried out to above-mentioned 2m+1 days based on average daily load are classified, and are found out and moment point place to be modified
The same type day of day;
B) to these same type days, construction includes time window composed by n before the amendment moment continuous moment points respectively, from
Two aspects of payload magnitude and load fluctuation trend, carry out and abnormal moment point institute this 2m time window curve respectively
In the similarity research of time window curve, it is based primarily upon the method for Euclidean distance similarity and the method based on cosine similarity,
Find at magnitude and the aspect of fluctuation tendency two with abnormal point where the more similar time window of time window curve as supporting
The training sample of vector machine SVM;Based on these samples, the load at moment where abnormal load data point is carried out based on SVM's
Fitting training, it is later that the data sequence before abnormal load data to be modified in time window is trained as being input to
In SVM model, the correction value of dominant abnormal data is obtained.
For given load data collectionConsideration is fitted with such as drag:
F (x)=(wx)+b (1)
Wherein, the number for the load similar day (the similar day of load form) that n is taken by load sample data set, yiFor mesh
Mark the load at time point, xiThe vector for being d for dimension, d time point corresponding load before numerical value is object time point
Amount;
W is real constant vector, w ∈ Rd, b is real constant, and the numerical value of b ∈ R, w and b are based on sample data set (xi,yi)n
Fitting obtains, and introduces the first slack variable ξiWith the second slack variableConstruct following optimization problem:
Wherein, constant C is penalty coefficient, and ε is prescribed skew value;
The optimization problem of its dual spaces are as follows:
Wherein, the first Lagrange multiplier αiWith the 2nd Lagrange multiplier αi *;
This problem is solved, the first optimal Lagrange multiplier α is obtainediWith the 2nd Lagrange multiplier αi *, to obtain
Fitting function
Wherein, K (x, xi)=(xxi+1)d, d is order, and to linear fit, d=1, gained fitting function f (x) is i.e. here
For the load correction value of required repairing.
(2) identification and amendment of recessive abnormal data
After dominant abnormal load data identification and amendment, needs to carry out daily load curve recessive abnormal data and distinguish
Know, since customer charge fluctuation has certain randomness, more abnormal data point is picked out from load fluctuation form
There is certain difficulty, it may appear that the case where crossing identification is considered as the method for wavelet decomposition and wavelet reconstruction to load curve
In rough error point recognized, and further compared by artificial or visual identity, to be confirmed whether being real abnormal number
According to.
Wavelet transformation can simultaneously analyze signal in time domain and frequency domain, can preferably distinguish making an uproar in signal
Sound, to realize the denoising to signal, the present invention is different to the recessiveness in customer charge sequence using the method for wavelet transformation
Often point is recognized.
Wavelet transformation is carried out to 96 point data of daily load first, chooses common wavelet basis db4, signal is carried out 4 layers points
Solution, the low frequency part after decomposition correspond to the trend part of signal, are the curve of a relative smooth, corresponding wavelet coefficient
For wk(k=1,2 ..., N);Rough error information is mainly reflected in the 1st layer of high frequency detail part d1 of signal decomposition, from radio-frequency head
Dividing can not can be clearly seen that the rough error point corresponding to signal finds the position of rough error, to height to preferably detect rough error point
Frequency part signal d1 carries out the denoising based on soft-threshold, and the selection of threshold value is directly related to denoising effect, chooses lesser
Threshold value can retain more wavelet coefficient, and the noise also retained simultaneously is also more;, whereas if the threshold value chosen is larger, retain
Noise with regard to less.
It is taken based on the weighted average threshold function table of soft and hard threshold function, i.e. semisoft shrinkage function, mathematic(al) representation are as follows:
Wherein, weighted factorT is threshold value thresholding, wkFor kth time wavelet coefficient, adopt
With the method for the VisuShrink of fixed threshold threshold criterion come threshold value thresholdingWherein σ is noise
Variance.
Further according to the wavelet coefficient after soft-threshold de-noisingWavelet reconstruction after de-noising is carried out to d1, it can be according to d1 de-noising
The extreme point of signal judges rough error position afterwards, as being likely to be the point of rough error caused by fluctuation situation, so need again into
The manual confirmation of one step is to find out really recessive exceptional value, if being confirmed as recessive exceptional value, takes dominant exceptional value amendment
Method be modified.
(2) the load point prediction model based on LSTM time recurrent neural network is established
Take the LSTM unit in LSTM Web vector graphic Fig. 1 as the node of hidden layer, LSTM unit specially devises note
Unit (memory cell) is recalled for saving historical information, the update and the utilization control by 3 doors respectively of historical information
System --- input gate (Input Gate) is forgotten door (Forget Gate), out gate (Output Gate).
If setting unit output h, input data x, g are LSTM unit output quantity, and i is the output quantity of input gate, and f is to forget door
Output, o are out gate output quantity, and c is memory unit output quantity, and h is entire LSTM unit output quantity, bg、bi、bf、boIt is ginseng
Base value is examined, the update of the unit of LSTM time recurrent neural network is made of following several formula:
gt=tanh (xtwxg+ht-1whg+bg) (6)
it=sigmoid (xtwxi+ht-1whi+ct-1wci+bi) (7)
ft=sigmoid (xtwxf+ht-1whf+ct-1wcf+bf) (8)
ot=sigmoid (xtwxo+ht-1who+bo) (9)
Dotted line connection in Fig. 1 is referred to as " peelhole connections ", 3 doors and independent memory
The design of cell, so that LSTM unit has preservation, reading, resetting and the ability for updating long range historical information.
Extracted from AMI system user over the past two years every 15 minutes load sequences, form raw data set, purpose
It is to do to predict a few days ago to 96 point data of daily load of user.
Consider that customer charge has apparent day morphological feature, 96 electrical load data of the previous day taken to be used as input,
It is sequentially sent to LSTM network, the reference point that next day 96 electrical load data are exported as network model ideal, therefore can be with
It determines, the input layer number l of LSTM network is 1, and output layer number of nodes O is 96.
The number of plies of hidden layer and its every layer of number of nodes have a significant impact the prediction effect of LSTM network model, here
Several structures as shown in Table 1 are rule of thumb selected, are filtered out by the effect of final interval prediction and data-oriented collection
More matched structure.
1 hidden layer structure of table
H in table 11Indicate the number of nodes of the first hidden layer, H2The number of nodes for indicating the second hidden layer, initial data is concentrated
All adjacent two days electrical load data are combined into a record, input of the previous day as LSTM network, one day after conduct
The label of network or output constitute data set D, the record for accounting for about 90% in data set D are enabled to account for about 10% as training set M
Record is used as test set N, and the record number of training set is m, and the record number of test set is n.
LSTM model is using the mean square error of predicted value and actual value as loss parameter, with the minimum optimization mesh of loss parameter
Mark carrys out each section weight of more new model, therefore, in order to judge the prediction of the LSTM network model with different implicit layer parameters
Effect defines the consensus forecast deviation of all moment points for test set:
Wherein real is the true value in test set sometime, and pred is the predicted value of its corresponding LSTM model, p=
96。
As it can be seen that Loss value is bigger, the deviation of predicted value and actual value is bigger, and prediction effect is also poorer;Loss value is smaller, in advance
Survey effect is also better, and there are the corresponding test set Loss data of LSTM structure institute of different implicit layer parameters to be shown in Table 2 institutes
Show.
The test set moment point consensus forecast deviation of 2 different parameters LSTM model of table
From table 2 it can be seen that the test set moment point consensus forecast deviation substantially difference of different parameters LSTM model is not
Greatly, emulation experiment is found, the average value of the power load data of test set is 126.20, in the enough situations of the number of iterations
Under, there is the LSTM network of different implicit layer parameters can preferably match data-oriented collection, and obtain similar in error result
Prediction result.
(3) the load setting prediction algorithm of point prediction value scaling coefficient is used
Interval prediction method is provided to predicting that the mobility scale of load quantifies as caused by uncertain factor
The forecast interval that Lower and upper bounds determine can be used for so that actual load observation is fallen in the section with certain expected probability
The load prediction of power-system short-term and ultra-short term.
In view of use the electro-mechanical wave situation difference at the different electricity consumption moment such as peak valley, using each moment as basic unit,
Interval prediction is carried out to each moment respectively.
(1) interval prediction evaluation index
Firstly, it is necessary to assess interval prediction result, the evaluation index used in interval prediction algorithm includes section
Coverage rate (Coverage Probability) χCP, mean breadth percentage (Mean Width Percentage) χMWPAnd accumulation
Deviation (Accumulated Deviation) χAD。
Section coverage rate χCPThe probability that actual value is fallen in the forecast interval by Lower and upper bounds envelope is defined, it is usually practical
Value realijIt is fallen in the probability not less than specified confidence level in constructed forecast interval, it may be assumed that
P(realij∈[L(predij),U(predij)])≥μ (13)
Wherein, L (predij) and U (predij) it is by point prediction value pred respectivelyijThe lower bound of obtained forecast interval and upper
Boundary, μ are that specified confidence interval is horizontal, and the corresponding section coverage rate of j-th of moment point is defined as follows:
Wherein,
Mean breadth percentage χMWPThe average percent that forecast interval width accounts for true value, j-th of moment point pair are measured
The mean breadth percentage answered is defined as follows:
It should be noted that using relative width rather than absolute width allows for, in load prediction problem, electricity consumption
Load when peak is usually big than other moment and it is difficult to predict the width of forecast interval ought to be wider, and when low power consumption
Prediction accuracy is relatively high, forecast interval relative narrower.
Other than section coverage rate and average width percentage two indices, there are one for the exterior point for falling in section
Evaluation index, i.e., on the case where giving specified confidence level, it is desirable to which those do not fall within point and prediction in forecast interval
The irrelevance in section is small as far as possible, in other words, in identical χCPAnd χMWPIn the case where, irrelevance is the smallest the result is that final choice
As a result, with cumulative departure χADTo embody the degree of deviation, the cumulative departure of j-th of moment point of definition are as follows:
Wherein,
(2) satisfaction of forecast interval
On the basis of giving specified confidence level μ, it is desirable to χCPjCan be as far as possible close to μ value, and χMWPjAnd χADjValue can be as far as possible
It is small, the overall target that quantitative evaluation can be carried out to forecast interval, i.e. forecast interval Satisfaction index (Prediction are proposed here
Interval Satisfaction Index)χPISI, the corresponding forecast interval Satisfaction index calculation formula of j-th of moment point
Are as follows:
Wherein, η is to χCPjThe penalty coefficient of value, sets according to actual needs, from formula (17) as can be seen that working as χCPjWith volume
Determine confidence level μ it is identical when, exponential term obtain minimum value 1.
With χCPjWith the increase of μ gap, exponential term is also increased rapidly, therefore, χPISIjValue it is smaller, corresponding section is pre-
It is more satisfactory to survey result, χPISIjThe upper bound corresponding to minimum value and lower bound are the interval prediction finally chosen as a result, in reality
In the calculating on border, it may appear that χMWPjAnd χCPjIt is identical, i.e. χPISIjIdentical situation, at this moment according to χADjValue is selected, will be minimum
χADjThe corresponding upper bound of value and lower bound are as the interval prediction result finally chosen.
(3) algorithmic descriptions
Using each moment as basic computational ele- ment, interval prediction is carried out to each moment respectively, for the j moment,
In training set, m predicted value and true Value Data correspondence are shared, predicted value is amplified respectively and reduces ratio factor alpha and β is obtained
The upper bound and lower bound of the result arrived as forecast interval, mathematic(al) representation are as follows:
Next it needs to be determined that the value of amplification and diminution ratio factor alpha and β, precision k refer to the precision of α and β, example
Such as, if the numerical value of precision k=0.01, α and β remain into 2 significant digits;If the numerical value of precision k=0.001, α and β are protected
Three are left to after decimal point;In view of the quantity of training set, take precision k=0.01 that satisfied effect can be obtained in actual calculating
Fruit.It enables:
α ' and β ' is the possibility value of α and β, and Ceil function is above to enter bracket function, such as Ceil (5.24)=6, Ceil
(5.00)=5, Ceil (- 5.24)=- 5, calculates entire training set, as a result duplicate removal, can obtain the value of α and β
The case where range, Fig. 2 is greater than predicted value with actual value, is illustrated algorithm.
For convenience, it is assumed that all predicted values are all identical and training set only has two groups of data, pred1j=as shown in Figure 2
Pred2j, value α and β corresponding to black dotted lines are respectively by real in figureij/predij- 1 and 1-realij/predijIt is calculated,
It is the position α ', β ' corresponding to solid black lines;In addition, differing precision k between adjacent two solid lines;First individually consider amplification
Proportionality coefficient α, in the case where lower bound determines, for real1jFor, it regard line 2 and line 3 as the upper bound, χCPjBe worth it is constant, still
Line 3 wants narrow as the forecast interval that the upper bound is constituted, χMWPjIt is worth small, therefore the case where line 2 is as the upper bound can directly exclude, line 3
Compared with line 1, the χ of line 3MWPjIt is worth small, line 1 is due to containing real2jIts χCPjValue is big, is both likely to be final choice
Compartmental results, therefore both of these case is involved in calculating, similarly, in the case where the upper bound determines, line 2 is compared with line 3 can be direct
The case where exclusion line 3, line 4 is compared with line 5, can directly exclude line 5, be less than predicted value for actual value, similarly meets above-mentioned
Therefore analysis can be obtained by the possibility value of all α and β by calculating α ' and β '.
α and β corresponding to forecast interval satisfaction minimum is chosen as the amplification of the moment point and reduces proportionality coefficient, it will
The coefficient is used in test set, obtains the upper bound and lower bound according to predicted value, passes through the relationship of actual value and the upper bound, lower bound, meter
The areal coverage and average width percentage for calculating test set, to verify interval prediction result.
(4) simulation case
1. large user's historical load data pre-processes
Simulating scenes: multiple large-scale power users, 2 years actual demand history data
Simulation result:
Fig. 3 (a) and (b) be respectively before certain user amendment and the revised data day of burden with power in 2 years curve, Fig. 4 (a) and
It (b) is user's daily load curve on the two, the moment 58,59,62 is missing point in Fig. 4 (a), black in revised value such as figure
Shown in color dot, the moment 64,89 is missing point in Fig. 4 (b), and revised value is as shown in black dot in figure.
Fig. 5 (a) and (b) are the revised daily load curve of user's missing values one day, missing values amendment black color dots mark
Out, wherein several days load curves in front and back of day where abnormal load point are also shown out to provide and compare by Fig. 5 (b).
2. the interval prediction based on LSTM point prediction
Simulating scenes: multiple large-scale power users, to its 2 years actual demand history data, with large user's historical load
What data prediction emulation obtained is training set to the data after anomalous data identification and repairing.
The amplification of each moment point is obtained with interval prediction algorithm on training set and reduces proportionality coefficient, in test set
The effect of upper verifying interval prediction model, enables:
In χCPIt is worth in identical situation, χMWPSmaller, then prediction result is better;In χMWPIt is worth in identical situation, χCPValue is got over
Both greatly, prediction result is better, therefore, comprehensively consider, construct evaluation index parameter lambda:
In χMWPAnd χCPIn similar situation, the effect of the larger interval prediction of λ is preferable, gives specified confidence level and is respectively
0.95, in the case where 0.90,0.85,0.80, the χ of the interval prediction model based on LSTM point prediction is calculatedCP、χMWPAnd λ.
Simulation result: having recorded in table 3 under every kind of confidence level, 5 LSTM hidden layer ginseng before evaluation index parameter lambda ranking
Several and its interval prediction is as a result, interval prediction result corresponding to all difference LSTM structures is shown in annex 1.
Interval prediction arithmetic result of the table 3 based on LSTM point prediction
From table 3 it is observed that interval prediction algorithm achieves good effect, average area coverage rate on test set
Slightly below given specified confidence level, but it is close enough with given specified confidence level, in addition, with specified
The reduction of confidence level, mean breadth percentages gradually decrease, and are consistent with expection, i.e., the corresponding area of narrower forecast interval
Between coverage rate it is then smaller, under four different specified confidence levels, two kinds of structures that Dan Yinhan and number of nodes are 100 and 300
Relatively good prediction result in the top is all obtained.
Fig. 6 is the predicted value and reality for some day that the LSTM model that containing single hidden layer and node in hidden layer is 100 obtains
Actual value comparison diagram, it can be seen from the figure that the result of point prediction can embody the trend of real data well, and error compared with
It is small.
Fig. 7 adds the interval prediction curve under four different confidence levels, Cong Tuzhong on the basis of Fig. 6 point prediction
As can be seen that actual value is largely fallen in the section of prediction, prediction effect is preferable, meanwhile, at the time of power load is low point with
Point compares at the time of power load is high, and the absolute width of forecast interval is narrower, this is also consistent with the setting of relative width.
3. the interval prediction algorithm comparison based on LSTM point prediction and NN point prediction
With the interval prediction based on LSTM point prediction under the same conditions, LSTM method and traditional neural network (NN)
Method comparing result is as follows:
Table 4LSTM and NN Comparative result
R_X is the relative error of index X, and the difference for corresponding to X index by NN and LSTM X index corresponding with LSTM is divided by
It arrives, negative value represents corresponding X index, and NN is less than LSTM, and positive value is opposite.
From table 4, it can be seen that LSTM has obvious advantage in all fields, average to miss compared with NN method
Difference is smaller, and section coverage rate is higher, and mean breadth percentage is smaller, and evaluation index parameter is more excellent.
Beneficial effect the present invention is based on the power consumer load setting prediction technique of deep learning is:
(1) accurate load setting prediction is carried out for the load of the single power consumer with strong stochastic volatility;
(2) it is different from the prediction technique for region load, and is substantially better than in the prediction accuracy of customer charge
Traditional method;
(3) electric power, the environmental protection and saving energy are provided to user with least cost under conditions of meeting safe and reliable.
The above description is only an embodiment of the present invention, is not intended to limit the scope of the invention, all to utilize this hair
Equivalent structure or equivalent flow shift made by bright description is applied directly or indirectly in other relevant technology necks
Domain is included within the scope of the present invention.
Claims (10)
1. a kind of power consumer load setting prediction technique based on deep learning, which comprises the following steps:
(1) identification and repairing of large user's historical load data pretreated model for abnormal data are established;
(2) the load point prediction model based on LSTM time recurrent neural network is established;
(3) the load setting prediction algorithm of point prediction value scaling coefficient is used.
2. the power consumer load setting prediction technique according to claim 1 based on deep learning, which is characterized in that institute
Stating abnormal data includes dominant abnormal data and recessive abnormal data.
3. the power consumer load setting prediction technique according to claim 2 based on deep learning, which is characterized in that institute
It states the identification of dominant abnormal data: checking that there are loads for any large user's load data discoveries in every 15 minutes in AMI system
It data the case where sometime point or certain continuous moment points lack, needs to fill up above-mentioned record with zero, and record
Abnormality mark;For already present data point in AMI system, the extreme abnormal conditions of numerical value will be present and pick out to come, and records different
Often mark;If the abnormality mark of above-mentioned record was present in k continuous moment, by the 96 point load values of this day from sample database
It is deleted, is otherwise repaired;
The repairing step of the dominant abnormal data:
A) it on the basis of day where moment point to be modified, respectively forwardly and backward finds m days, if the m days in the future load points have
Dominant abnormality mark and it is uncorrected, postpone backward one day;It finds out respectively 2m days and the average daily of moment point to be modified place day is born
Lotus, it is contemplated that the difference on working day and day off classifies to the similar day carried out based on average daily load for above-mentioned 2m+1 days, find out with
The same type day of day where moment point to be modified;
B) same type day is constructed respectively comprising time window composed by n before the amendment moment continuous moment points, from payload
Two aspects of magnitude and load fluctuation trend carry out and time window where abnormal moment point this 2m time window curve respectively
The similarity research of curve, is based primarily upon the method for Euclidean distance similarity and the method based on cosine similarity, and searching is being measured
Two aspects of value and fluctuation tendency time window more similar with time window curve where abnormal point is as support vector machines
Training sample.Based on these samples, the training of the fitting based on SVM is carried out to the load at moment where abnormal load data point,
Later using the data sequence before abnormal load data to be modified in time window as being input in trained SVM model,
Obtain the correction value of dominant abnormal data;
For given load data collectionxi∈Rd,yi∈ R is fitted with such as drag:
F (x)=(wx)+b (1)
Wherein, the number for the load similar day (the similar day of load form) that wherein n is taken by load sample data set, yiFor mesh
Mark the load at time point, xiThe vector for being d for dimension, d time point corresponding load before numerical value is object time point
Amount;
W is real constant vector, w ∈ Rd, b is real constant, and the numerical value of b ∈ R, w and b are based on sample data set (xi,yi)nFitting
It obtains, introduces the first slack variable ξiWith the second slack variableConstruct following optimization problem:
Wherein, constant C is penalty coefficient, and ε is prescribed skew value;
The optimization problem of its dual spaces are as follows:
Wherein, the first Lagrange multiplier αiWith the 2nd Lagrange multiplier αi *;
This problem is solved, the first optimal Lagrange multiplier α is obtainediWith the 2nd Lagrange multiplier αi *, to be fitted
Function
Wherein, K (x, xi)=(xxi+ 1) the load correction value repaired needed for, gained fitting function f (x) is here.
4. the power consumer load setting prediction technique according to claim 2 based on deep learning, which is characterized in that adopt
The step of recessive abnormal data in customer charge sequence is recognized with wavelet decomposition and the method for wavelet reconstruction: first
Wavelet transformation first is carried out to 96 point data of daily load, chooses wavelet basis db4, signal is subjected to 4 layers of decomposition, the low frequency portion after decomposition
Dividing the trend part corresponding to signal is the curve of a relative smooth, and corresponding wavelet coefficient is wk(k=1,2 ..., N),
Denoising based on soft-threshold is carried out to high frequency section signal, is taken based on the weighted average threshold value letter of soft and hard threshold function
Number:
Wherein, weighted factorT is threshold value thresholding, wkFor kth time wavelet coefficient, adopt
With the method for the VisuShrink of fixed threshold threshold criterion come threshold value thresholdingWherein σ is the side of noise
Difference, further according to the wavelet coefficient after soft-threshold de-noisingWavelet reconstruction after de-noising is carried out to high frequency section signal, according to high frequency
The extreme point of signal judges rough error position after part signal de-noising, and excludes the point of the rough error as caused by fluctuation situation,
If being confirmed as recessive exceptional value, the modified method of dominant exceptional value is taken to be modified.
5. the power consumer load setting prediction technique according to claim 1 based on deep learning, which is characterized in that root
According to large user's historical load data pretreated model, extracted from power information acquisition system user every 15 minutes interval
Historical load sequence forms raw data set, does prediction a few days ago to 96 point data of daily load of user and obtains customer charge prediction
Value;It is the training set of neural network based on actual power consumer load data, test obtains optimal time recurrent neural net
Network LSTM hidden layer configuration;Using the mean square error of customer charge predicted value and actual value as loss parameter, with loss parameter minimum
Each section weight of LSTM model is updated for optimization aim, obtains optimum load prediction model.
6. the power consumer load setting prediction technique according to claim 5 based on deep learning, which is characterized in that institute
Stating time recurrent neural network LSTM hidden layer configuration includes:
If unit exports h, input data x, g are LSTM unit output quantity, and i is the output quantity of input gate, and f is to forget door output, o
For out gate output quantity, c is memory unit output quantity, and h is entire LSTM unit output quantity, bg、bi、bf、boIt is with reference to base
Value, the update of the unit of LSTM time recurrent neural network are as follows:
gt=tanh (xtwxg+ht-1whg+bg) (6)
it=sigmoid (xtwxi+ht-1whi+ct-1wci+bi) (7)
ft=sigmoid (xtwxf+ht-1whf+ct-1wcf+bf) (8)
ot=sigmoid (xtwxo+ht-1who+bo) (9)
Extracted from power information acquisition system user nearly 2 years every 15 minutes load sequences, form raw data set,
All adjacent two days electrical load data are concentrated to be combined into a record initial data, the previous day is defeated as LSTM network
Enter, one day after as the label of network or output, constitutes data set D, enable the record for accounting for 90% in data set D as training set
M accounts for 10% record as test set N, and the record number of training set is m, and the record number of test set is n;
LSTM model is come using the mean square error of predicted value and actual value as loss parameter with the minimum optimization aim of loss parameter
Each section weight of more new model defines the consensus forecast deviation of all moment points:
Wherein real is the true value in test set sometime, and pred is the predicted value of its corresponding LSTM model, p=96.
7. the power consumer load setting prediction technique according to claim 1 based on deep learning, which is characterized in that root
According to the load point prediction model to predicting that the mobility scale of load quantifies as caused by uncertain factor, provide
The forecast interval that Lower and upper bounds determine, so that actual load observation is fallen in the section with certain expected probability.
8. the power consumer load setting prediction technique according to claim 7 based on deep learning, which is characterized in that institute
The evaluation index for stating the variation range of the prediction load of forecast interval includes section coverage rate χCP, mean breadth percentage χMWPWith
Cumulative departure χAD;
Section coverage rate χCPBe actual value fall in by the upper bound, lower bound envelope forecast interval in probability, actual value realijWith not
Probability lower than specified confidence level is fallen in the forecast interval of building, it may be assumed that
P(realij∈[L(predij),U(predij)])≥μ (13)
Wherein, L (predij) and U (predij) it is by point prediction value pred respectivelyijThe lower bound of obtained forecast interval and the upper bound, μ
Horizontal for specified confidence interval, the corresponding section coverage rate of j-th of moment point is defined as follows:
Wherein,
Mean breadth percentage χMWPThe average percent that forecast interval width accounts for true value is measured, j-th of moment point is corresponding
Mean breadth percentage is defined as follows:
Cumulative departure χADTo embody the degree that actual load observation deviates forecast interval, the cumulative departure of j-th of moment point are as follows:
Wherein,
9. the power consumer load setting prediction technique according to claim 7 based on deep learning, which is characterized in that institute
State the Satisfaction index χ of forecast intervalPISIIn j-th of moment point are as follows:
Wherein, η is to χCPjThe penalty coefficient of value, χPISIjThe upper bound corresponding to minimum value and lower bound are the section finally chosen
Prediction result works as χMWPjAnd χCPjWhen identical, according to χADjValue is selected, by the smallest χADjThe corresponding upper bound of value and lower bound
As the interval prediction result finally chosen.
10. the power consumer load setting prediction technique according to claim 1 based on deep learning, which is characterized in that
The load setting prediction algorithm carries out interval prediction to each moment respectively, when j using each moment as basic computational ele- ment
It is engraved in training set shared m predicted value and true Value Data is corresponding, predicted value is amplified to respectively and reduced ratio factor alpha and β
The upper bound and lower bound of the obtained result as forecast interval:
It determining amplification and reduces the value of ratio factor alpha and β, precision k is the precision of α and β, it enables:
α ' and β ' is the possibility value of α and β, and Ceil function is above to enter bracket function, carries out calculated result to entire training set and goes
Again, the value range of α and β can be obtained, chooses α and β corresponding when the Satisfaction index minimum of forecast interval as this
The amplification of moment point and diminution proportionality coefficient, which is used in test set, obtains the upper bound and lower bound according to predicted value, is led to
The relationship for crossing actual value and the upper bound, lower bound calculates the areal coverage and average width percentage of test set, verifies interval prediction
As a result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910550680.7A CN110263866B (en) | 2019-06-24 | 2019-06-24 | Power consumer load interval prediction method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910550680.7A CN110263866B (en) | 2019-06-24 | 2019-06-24 | Power consumer load interval prediction method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110263866A true CN110263866A (en) | 2019-09-20 |
CN110263866B CN110263866B (en) | 2023-11-10 |
Family
ID=67920963
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910550680.7A Active CN110263866B (en) | 2019-06-24 | 2019-06-24 | Power consumer load interval prediction method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110263866B (en) |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110674993A (en) * | 2019-09-26 | 2020-01-10 | 广东电网有限责任公司 | User load short-term prediction method and device |
CN110991761A (en) * | 2019-12-13 | 2020-04-10 | 新奥数能科技有限公司 | Heat supply load prediction method and device |
CN111090684A (en) * | 2019-12-17 | 2020-05-01 | 国网电子商务有限公司 | Method and system for processing electricity consumption data |
CN111222689A (en) * | 2019-12-06 | 2020-06-02 | 重庆三峡学院 | LSTM load prediction method, medium, and electronic device based on multi-scale temporal features |
CN111311436A (en) * | 2020-03-17 | 2020-06-19 | 上海积成能源科技有限公司 | System and method for analyzing and managing comprehensive energy consumption of complex renewable energy sources based on multivariable control |
CN111325413A (en) * | 2020-03-23 | 2020-06-23 | 电子科技大学 | Short-term load interval prediction method based on neural network and proportionality coefficient method |
CN111461462A (en) * | 2020-04-29 | 2020-07-28 | 南京工程学院 | Daily load prediction method based on TrellisNet-L STM |
CN111582542A (en) * | 2020-03-31 | 2020-08-25 | 国网上海市电力公司 | Power load prediction method and system based on abnormal restoration |
CN111598330A (en) * | 2020-05-13 | 2020-08-28 | 国网河北省电力有限公司电力科学研究院 | Residential electricity load prediction method based on load identification technology |
CN111666276A (en) * | 2020-06-11 | 2020-09-15 | 上海积成能源科技有限公司 | Method for eliminating abnormal data by applying isolated forest algorithm in power load prediction |
CN111784028A (en) * | 2020-06-08 | 2020-10-16 | 深圳供电局有限公司 | Community load prediction method |
CN111784019A (en) * | 2019-12-26 | 2020-10-16 | 国网北京市电力公司 | Power load processing method and device |
CN112070238A (en) * | 2020-11-10 | 2020-12-11 | 鹏城实验室 | Accurate machine learning asynchronous prediction method and system and storage medium |
CN112183813A (en) * | 2020-08-26 | 2021-01-05 | 河海大学 | Ultra-short-term load rolling multi-step prediction method based on optimized sparse coding |
CN112232714A (en) * | 2020-11-18 | 2021-01-15 | 中国科学院电工研究所 | Power distribution network risk assessment method under incomplete structural parameters based on deep learning |
CN112653241A (en) * | 2020-12-16 | 2021-04-13 | 南方电网深圳数字电网研究院有限公司 | Power grid monitoring method and device and Internet of things metering unit |
CN112734141A (en) * | 2021-02-25 | 2021-04-30 | 云南电网有限责任公司电力科学研究院 | Diversified load interval prediction method and device |
CN112990587A (en) * | 2021-03-24 | 2021-06-18 | 北京市腾河智慧能源科技有限公司 | Method, system, equipment and medium for accurately predicting power consumption of transformer area |
CN113139605A (en) * | 2021-04-27 | 2021-07-20 | 武汉理工大学 | Power load prediction method based on principal component analysis and LSTM neural network |
CN114219150A (en) * | 2021-12-15 | 2022-03-22 | 浙江大学 | Power load interval prediction method based on self-adaptive optimization construction interval |
CN115630830A (en) * | 2022-12-01 | 2023-01-20 | 北京忠业兴达科技有限公司 | Power supply and distribution method, device, equipment and storage medium for data center |
CN116361709A (en) * | 2023-03-31 | 2023-06-30 | 山东省计算中心(国家超级计算济南中心) | Self-adaptive power load identification method and device |
CN116470618A (en) * | 2023-04-17 | 2023-07-21 | 深圳市威能讯电子有限公司 | Mobile outdoor energy storage charge and discharge control method |
CN117349778A (en) * | 2023-12-04 | 2024-01-05 | 湖南蓝绿光电科技有限公司 | Online real-time monitoring system of consumer based on thing networking |
CN117472898A (en) * | 2023-12-26 | 2024-01-30 | 国网江西省电力有限公司电力科学研究院 | Fusion-based power distribution network abnormal data error correction method and system |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106960252A (en) * | 2017-03-08 | 2017-07-18 | 深圳市景程信息科技有限公司 | Methods of electric load forecasting based on long Memory Neural Networks in short-term |
CN109376772A (en) * | 2018-09-28 | 2019-02-22 | 武汉华喻燃能工程技术有限公司 | A kind of Combination power load forecasting method based on neural network model |
CN109376960A (en) * | 2018-12-06 | 2019-02-22 | 国网北京市电力公司 | Load Forecasting based on LSTM neural network |
-
2019
- 2019-06-24 CN CN201910550680.7A patent/CN110263866B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106960252A (en) * | 2017-03-08 | 2017-07-18 | 深圳市景程信息科技有限公司 | Methods of electric load forecasting based on long Memory Neural Networks in short-term |
CN109376772A (en) * | 2018-09-28 | 2019-02-22 | 武汉华喻燃能工程技术有限公司 | A kind of Combination power load forecasting method based on neural network model |
CN109376960A (en) * | 2018-12-06 | 2019-02-22 | 国网北京市电力公司 | Load Forecasting based on LSTM neural network |
Non-Patent Citations (4)
Title |
---|
于佳弘等: "基于LSTM 的用户负荷区间预测方法", 工业控制计算机, vol. 31, no. 4, pages 100 - 102 * |
于佳弘等: "基于LSTM的用户负荷区间预测方法", 《工业控制计算机》 * |
于佳弘等: "基于LSTM的用户负荷区间预测方法", 《工业控制计算机》, vol. 31, no. 04, 25 April 2018 (2018-04-25), pages 100 - 102 * |
王国玲: "电力系统短期负荷预测方法的研究", 中国优秀硕士学位论文全文数据库工程科技Ⅱ辑(月刊), no. 4, pages 264 - 265 * |
Cited By (40)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110674993A (en) * | 2019-09-26 | 2020-01-10 | 广东电网有限责任公司 | User load short-term prediction method and device |
CN111222689A (en) * | 2019-12-06 | 2020-06-02 | 重庆三峡学院 | LSTM load prediction method, medium, and electronic device based on multi-scale temporal features |
CN110991761A (en) * | 2019-12-13 | 2020-04-10 | 新奥数能科技有限公司 | Heat supply load prediction method and device |
CN110991761B (en) * | 2019-12-13 | 2022-06-28 | 新奥数能科技有限公司 | Heat supply load prediction method and device |
CN111090684A (en) * | 2019-12-17 | 2020-05-01 | 国网电子商务有限公司 | Method and system for processing electricity consumption data |
CN111090684B (en) * | 2019-12-17 | 2023-11-03 | 国网数字科技控股有限公司 | Method and system for processing electricity consumption data |
CN111784019A (en) * | 2019-12-26 | 2020-10-16 | 国网北京市电力公司 | Power load processing method and device |
CN111311436A (en) * | 2020-03-17 | 2020-06-19 | 上海积成能源科技有限公司 | System and method for analyzing and managing comprehensive energy consumption of complex renewable energy sources based on multivariable control |
CN111325413A (en) * | 2020-03-23 | 2020-06-23 | 电子科技大学 | Short-term load interval prediction method based on neural network and proportionality coefficient method |
CN111325413B (en) * | 2020-03-23 | 2022-04-01 | 电子科技大学 | Short-term load interval prediction method based on neural network and proportionality coefficient method |
CN111582542A (en) * | 2020-03-31 | 2020-08-25 | 国网上海市电力公司 | Power load prediction method and system based on abnormal restoration |
CN111582542B (en) * | 2020-03-31 | 2023-10-03 | 国网上海市电力公司 | Power load prediction method and system based on anomaly repair |
CN111461462A (en) * | 2020-04-29 | 2020-07-28 | 南京工程学院 | Daily load prediction method based on TrellisNet-L STM |
CN111461462B (en) * | 2020-04-29 | 2023-11-21 | 南京工程学院 | Daily load prediction method based on trellis Net-LSTM |
CN111598330A (en) * | 2020-05-13 | 2020-08-28 | 国网河北省电力有限公司电力科学研究院 | Residential electricity load prediction method based on load identification technology |
CN111784028A (en) * | 2020-06-08 | 2020-10-16 | 深圳供电局有限公司 | Community load prediction method |
CN111666276A (en) * | 2020-06-11 | 2020-09-15 | 上海积成能源科技有限公司 | Method for eliminating abnormal data by applying isolated forest algorithm in power load prediction |
CN112183813B (en) * | 2020-08-26 | 2024-04-09 | 河海大学 | Ultra-short-term load rolling multi-step prediction method based on optimized sparse coding |
CN112183813A (en) * | 2020-08-26 | 2021-01-05 | 河海大学 | Ultra-short-term load rolling multi-step prediction method based on optimized sparse coding |
CN112070238B (en) * | 2020-11-10 | 2021-02-23 | 鹏城实验室 | Accurate machine learning asynchronous prediction method and system and storage medium |
CN112070238A (en) * | 2020-11-10 | 2020-12-11 | 鹏城实验室 | Accurate machine learning asynchronous prediction method and system and storage medium |
CN112232714A (en) * | 2020-11-18 | 2021-01-15 | 中国科学院电工研究所 | Power distribution network risk assessment method under incomplete structural parameters based on deep learning |
CN112232714B (en) * | 2020-11-18 | 2023-06-20 | 中国科学院电工研究所 | Deep learning-based risk assessment method for distribution network under incomplete structural parameters |
CN112653241A (en) * | 2020-12-16 | 2021-04-13 | 南方电网深圳数字电网研究院有限公司 | Power grid monitoring method and device and Internet of things metering unit |
CN112653241B (en) * | 2020-12-16 | 2023-06-09 | 南方电网数字平台科技(广东)有限公司 | Power grid monitoring method and device and Internet of things metering unit |
CN112734141A (en) * | 2021-02-25 | 2021-04-30 | 云南电网有限责任公司电力科学研究院 | Diversified load interval prediction method and device |
CN112734141B (en) * | 2021-02-25 | 2023-01-20 | 云南电网有限责任公司电力科学研究院 | Diversified load interval prediction method and device |
CN112990587B (en) * | 2021-03-24 | 2023-10-24 | 北京市腾河智慧能源科技有限公司 | Method, system, equipment and medium for accurately predicting power consumption of transformer area |
CN112990587A (en) * | 2021-03-24 | 2021-06-18 | 北京市腾河智慧能源科技有限公司 | Method, system, equipment and medium for accurately predicting power consumption of transformer area |
CN113139605A (en) * | 2021-04-27 | 2021-07-20 | 武汉理工大学 | Power load prediction method based on principal component analysis and LSTM neural network |
CN114219150A (en) * | 2021-12-15 | 2022-03-22 | 浙江大学 | Power load interval prediction method based on self-adaptive optimization construction interval |
CN115630830A (en) * | 2022-12-01 | 2023-01-20 | 北京忠业兴达科技有限公司 | Power supply and distribution method, device, equipment and storage medium for data center |
CN116361709A (en) * | 2023-03-31 | 2023-06-30 | 山东省计算中心(国家超级计算济南中心) | Self-adaptive power load identification method and device |
CN116361709B (en) * | 2023-03-31 | 2023-10-31 | 山东省计算中心(国家超级计算济南中心) | Self-adaptive power load identification method and device |
CN116470618A (en) * | 2023-04-17 | 2023-07-21 | 深圳市威能讯电子有限公司 | Mobile outdoor energy storage charge and discharge control method |
CN116470618B (en) * | 2023-04-17 | 2024-01-26 | 深圳市威能讯电子有限公司 | Mobile outdoor energy storage charge and discharge control method |
CN117349778A (en) * | 2023-12-04 | 2024-01-05 | 湖南蓝绿光电科技有限公司 | Online real-time monitoring system of consumer based on thing networking |
CN117349778B (en) * | 2023-12-04 | 2024-02-20 | 湖南蓝绿光电科技有限公司 | Online real-time monitoring system of consumer based on thing networking |
CN117472898A (en) * | 2023-12-26 | 2024-01-30 | 国网江西省电力有限公司电力科学研究院 | Fusion-based power distribution network abnormal data error correction method and system |
CN117472898B (en) * | 2023-12-26 | 2024-04-02 | 国网江西省电力有限公司电力科学研究院 | Fusion-based power distribution network abnormal data error correction method and system |
Also Published As
Publication number | Publication date |
---|---|
CN110263866B (en) | 2023-11-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110263866A (en) | A kind of power consumer load setting prediction technique based on deep learning | |
CN112508275B (en) | Power distribution network line load prediction method and equipment based on clustering and trend indexes | |
CN113962364B (en) | Multi-factor power load prediction method based on deep learning | |
CN110097297B (en) | Multi-dimensional electricity stealing situation intelligent sensing method, system, equipment and medium | |
De Silva et al. | A data mining framework for electricity consumption analysis from meter data | |
CN108388962B (en) | Wind power prediction system and method | |
Enke et al. | A hybrid neuro-fuzzy model to forecast inflation | |
CN110414788A (en) | A kind of power quality prediction technique based on similar day and improvement LSTM | |
CN111401599B (en) | Water level prediction method based on similarity search and LSTM neural network | |
Zhang et al. | Power consumption predicting and anomaly detection based on transformer and K-means | |
CN112329990A (en) | User power load prediction method based on LSTM-BP neural network | |
CN107730059A (en) | The method of transformer station's electricity trend prediction analysis based on machine learning | |
CN112215426A (en) | Short-term power load prediction method | |
CN111178585A (en) | Fault reporting amount prediction method based on multi-algorithm model fusion | |
CN106530082A (en) | Stock predication method and stock predication system based on multi-machine learning | |
Wang et al. | A novel decomposition-ensemble forecasting system for dynamic dispatching of smart grid with sub-model selection and intelligent optimization | |
CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
CN112949207A (en) | Short-term load prediction method based on improved least square support vector machine | |
CN115860177A (en) | Photovoltaic power generation power prediction method based on combined machine learning model and application thereof | |
CN116169670A (en) | Short-term non-resident load prediction method and system based on improved neural network | |
CN108229750A (en) | A kind of stock yield Forecasting Methodology | |
CN105488598A (en) | Medium-and-long time electric power load prediction method based on fuzzy clustering | |
CN110188956A (en) | Load forecasting method based on ant colony neural network and the system comprising this method, memory | |
CN108345996B (en) | System and method for reducing wind power assessment electric quantity | |
Gao et al. | Establishment of economic forecasting model of high-tech industry based on genetic optimization neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |